Effects of class purity of training data on crop classification using 2D-CNn

초록

Proper collection of training data is an important factor in supervised classification for crop mapping. Each pixel in remote sensing imagery represents an area with various characteristics of surface objects and may have different spectral values for the same crop type. This mixed pixel effect in training data may greatly affect classification results. Although much effort has been made for the proper selection of training data in pixel-based classification, few studies have been conducted in patch-based classification with deep learning. In this study, we analyze the effect of class purity within the patch of training data on a patch-based 2D convolutional neural network (2D-CNN) model for crop classification. The classification performance of 2D-CNN was evaluated from two case study areas with different spatial characteristics of crops and input images with different spatial resolutions. In the area which consists of crop parcels with similar shapes and uniform patterns, the classification accuracy could be improved by collecting training samples with high class purity in the high spatial resolution imagery. On the contrary, using training samples with lower class purity in crop classification with Landsat images led to the improvement in the classification accuracy in the classification of areas where crop parcels had various shapes and sizes. These experimental results indicate that training data in the patch-based crop classification should be selected by taking into account the characteristics of the area to be classified. © 2020 40th Asian Conference on Remote Sensing, ACRS 2019: Progress of Remote Sensing Technology for Smart Future. All rights reserved.

키워드